Can Large Audio-Language Models Ignore Multilingual Distractors? An Evaluation of Their Selective Auditory Attention Capabilities

A study by Heejoon Koo, published in the Proceedings of the IMPACT-SPEECH Workshop (pages 20–32, Budapest, October 2026), introduced MUSI, a cocktail-party-inspired diagnostic benchmark pairing an English target dialogue with a plausible distractor in English, Spanish, Korean, or Chinese. Testing four open-weight and two closed-source large audio-language models across single-stream, separation-based, and end-to-end cocktail party settings at controlled SNRs, Koo found model-dependent language-conditioned variation and heightened vulnerability at adverse SNRs, with errors dominated by distractor-grounded source confusion. Separation reduced acoustic overlap but often left source attribution unresolved, leading Koo to identify selective auditory attention as a primary capability for reliable LALM deployment.

Abstract Robust selective auditory attention under multilingual interference is critical for reliable LALM deployment. We introduce MUSI, a cocktail party-inspired controlled diagnostic evaluation of source-grounded spokenlanguage understanding and reasoning. Each item pairs an English target dialogue with a plausible distractor in English, Spanish, Korean, or Chinese, and evaluates models under 1 single-stream, 2 separation-based, and 3 end-to-end cocktail party settings across controlled SNRs. Across four open-weight and two closed-source LALMs, we find modeldependent language-conditioned variation and heightened vulnerability at adverse SNRs. Errors are dominated by distractor-grounded source confusion, while separation reduces acoustic overlap but often leaves source attribution unresolved. These findings highlight selective auditory attention as a primary capability for reliable LALM deployment. - Anthology ID: - 2026.impactspeech-1.3 - Volume: - Proceedings of the IMPACT-SPEECH Workshop: Identifying, Measuring, Preventing, and Assessing Consequences of Bias in Speech LLMs https://aclanthology.org/volumes/2026.impactspeech-1/ - Month: - October - Year: - 2026 - Address: - Budapest, Hungary - Editors: - Ravi Shekhar https://aclanthology.org/people/ravi-shekhar/ , Monorama Swain https://aclanthology.org/people/monorama-swain/ , Jagabandhu Mishra https://aclanthology.org/people/jagabandhu-mishra/unverified/ , Sandipan Dhar https://aclanthology.org/people/sandipan-dhar/unverified/ , Haralambos Mouratidis https://aclanthology.org/people/haralambos-mouratidis/ , Matthew Purver https://aclanthology.org/people/matthew-purver/ - Venues: - IMPACT-SPEECH https://aclanthology.org/venues/impactspeech/ | WS https://aclanthology.org/venues/ws/ - SIG: - Publisher: - Association for Computational Linguistics - Note: - Pages: - 20–32 - Language: - URL: - https://aclanthology.org/2026.impactspeech-1.3/ https://aclanthology.org/2026.impactspeech-1.3/ - DOI: - Cite ACL : - Heejoon Koo. 2026. Can Large Audio-Language Models Ignore Multilingual Distractors? An Evaluation of Their Selective Auditory Attention Capabilities https://aclanthology.org/2026.impactspeech-1.3/ . In Proceedings of the IMPACT-SPEECH Workshop: Identifying, Measuring, Preventing, and Assessing Consequences of Bias in Speech LLMs , pages 20–32, Budapest, Hungary. Association for Computational Linguistics. - Cite Informal : - Can Large Audio-Language Models Ignore Multilingual Distractors? An Evaluation of Their Selective Auditory Attention Capabilities https://aclanthology.org/2026.impactspeech-1.3/ Koo, IMPACT-SPEECH 2026 - PDF: - https://aclanthology.org/2026.impactspeech-1.3.pdf https://aclanthology.org/2026.impactspeech-1.3.pdf